| abstract |
A pedestrian re-identification method based on large-interval relative distance metric learning, comprising the steps of: performing dimensionality reduction processing on the feature expression vectors of pedestrian images in the training data set and further projecting the dimensionality-reduced vectors into intra-class subspaces; according to the projection The final feature expression vector and corresponding label information are used to learn the Mahalanobis distance metric matrix by optimizing the loss function; in the test data set, the learned Mahalanobis distance metric matrix is used to perform pedestrian re-identification on pedestrian images under different cameras. Since the Mahalanobis distance metric matrix is learned through relative distance comparison in the intra-class subspace, the obtained metric matrix is more robust, and the accuracy of pedestrian re-identification is significantly improved on the test set. |